Haas O (2024)
Publication Language: English
Publication Type: Thesis
Publication year: 2024
URI: https://open.fau.de/handle/openfau/31430
Clinics and hospitals are complex entities that fill different roles. This is reflected by heterogeneous clinical data, which come from different data sources and domains like radiology, controlling, or medicine, but all describe patients and their care. This thesis presents novel methods that address challenges that arise when using heterogeneous clinical data and discusses what opportunities arise from their use. Four challenges have been identified in this thesis. First, incompatible systems require manual work if data from different systems have to be combined. Second, multi-model data and tasks increase the complexity. Third, a multitude of variables can render methods infeasible or less efficient. Lastly, lacking interpretability hinders a human's understanding of models derived from the data. This thesis discusses the use of Business Intelligence and Machine Learning methods. Four contributions are discussed in this thesis. First, the use of Business Intelligence methods for the automation of nursing management documentation in the presence of incompatible systems is developed. This allows managers to speed up this otherwise time-consuming and error-prone process. Novel interpretable, rule-based Machine Learning methods have been developed to solve three tasks which are multi-modal, contain multi-modal data, and are prone to lacking interpretability. In the first task, standardized questionnaire data is used to predict anxiety in palliative care patients with a novel method based on the Association Rule Mining metrics lift and local support. The model allows providers to assess which patients might suffer from anxiety in order to provide treatment. The other two tasks are also affected by a multitude of variables in the dataset. In the first of these, the risk for in-hospital mortality is estimated with a novel method using the epidemiological metric odds ratio. The predictive model allows providers to assess a patient's in-hospital mortality risk and allocate resources adequately. In the second contribution, a patient's previous admissions to the hospital are used in a novel method that builds upon the epidemiological metric incidence rate ratios to estimate their risk of being infected with the human immunodeficiency virus in the future. This estimated risk can serve as a first risk assessment prior to testing or pre-exposure prophylaxis counseling. These contributions lead to the following opportunities. The Business Intelligence methods employed in this thesis allow nursing managers to get additional insight via visualization capabilities, which facilitate the analysis of already existing data. The Machine Learning methods, on the other hand, provide high predictive performance and can be used with other tasks, including tasks in domains other than healthcare. Additionally, the interpretable nature of the models allows researchers to analyze the correlations between the variables in heterogeneous clinical data.
APA:
Haas, O. (2024). Heterogeneous Clinical Data - Challenges and Opportunities (Dissertation).
MLA:
Haas, Oliver. Heterogeneous Clinical Data - Challenges and Opportunities. Dissertation, 2024.
BibTeX: Download